Customers know the style they want when they see it — they just can't search for it. How multimodal retrieval plus an LLM layer turned "make my room feel like this" into a shoppable experience, and why the key decision was choosing what the model was allowed to be wrong about.

Most personalization quietly ignores anonymous visitors — the majority of traffic. The story of a browse re-ranker that treated cold start as the main event, not an edge case, and drove tens of millions in incremental revenue.

A central recs team had become the company's experimentation bottleneck. Instead of hiring faster, we discovered that hundreds of "unique" internal requests were the same five primitives in disguise — and built a self-service platform around them.

Every AI product stands on a data layer someone has to product-manage. Cutting producer onboarding from days to minutes, making failed events debuggable instead of invisible, and finding multi-million-dollar savings in the cloud bill.

Before ML personalization: firmware for vehicles you can't hotfix, and data science features for small-business owners with no analytics team. Why range compounds instead of resetting.

You can't A/B test your way out of hallucination. A practical framework for evaluating LLM-powered experiences — relevance, coherence, brand safety, taste — written for PMs who need to define "good" before engineering can ship it.